Anthropic has begun implementing an invisible, machine-readable watermark in text generated by supported Claude models worldwide. Unlike metadata or hidden characters, as I understand it, the watermark arises from statistical patterns in Claude’s word choices. For lawyers, the critical point is that the watermark signals processing, not authorship. A brief drafted entirely by Claude may carry the same type of signature as a lawyer-written brief sent to Claude solely for grammar, style, translation, or proofreading. Because watermarking occurs at the model level, it applies across Claude-supported products and integrations, and the signal travels with copied text and may survive some editing. A detected watermark therefore does not establish who developed the argument, conducted the research, or exercised legal judgment.

That distinction matters as courts adopt a patchwork of rules governing generative AI. Some judges require disclosure whenever AI assists in preparing a filing; others require certification that AI-generated material, authorities, and factual assertions were independently verified; still others rely on existing obligations under Rule 11. Anthropic has announced plans to provide a detection API and support third-party detection, potentially enabling courts, clerks, and opposing counsel to test filings for Claude’s signature. Detection, however, will be probabilistic rather than conclusive, while heavy editing, short passages, or mixed-source text may weaken or eliminate the signal.

Watermarking could become an audit tool, but not an authorship test. A court might use a detected mark to scrutinize whether counsel complied with an applicable disclosure or certification rule, even when Claude’s role was limited to polishing human work. The prudent response is not to avoid responsible AI use, but to strengthen governance: identify applicable judge-specific rules, document how AI was used, preserve meaningful human review, verify every factual and legal proposition, and protect client information. These practices align with lawyers’ continuing duties of competence, confidentiality, supervision, candor, and independent verification.

Lawyers should understand not only what an AI tool produces, but also what signals it leaves behind—and what those signals do, and do not, prove.